[visarga]

They can benchmark a few architectures (ResNet50, BERT) and tell us how many times we can train a model on a specific level of subscription.